A Comparative Analysis of Machine Learning Algorithms for Breast Cancer Detection and Identification of Key Predictive Features

Breast cancer, which develops from the cells of the breast, is a prevalent health concern. Considerable funding for breast cancer awareness and research has fueled progress in the recognition and treatment of this disease, effectively reducing mortality rates. Thanks to such advancements, coupled with earlier detection, innovative treatment options, and a more comprehensive understanding of the disease, survival rates have improved and associated deaths have declined. Within the scope of this research, our focus narrows to recognize breast cancer using machine learning.

The concept of classification, the process of sorting feature vectors into more than one classes, can vary among algorithms, and choosing an appropriate algorithm for a certain task involves many decisions. A primary challenge in the early identification of breast cancer lies in distinguishing benign from malignant tumors. Benign breast cancers are characterized by noncancerous cells, while malignant ones consist of cancerous cells that may start to spread. Early intervention can prevent benign tumors from progressing to dangerous stages.

To efficiently categorize breast cancer as benign or malignant, we propose a fast-acting predictive model for cancer categorization, operating within a human-computer interactive (HCI) diagnosis framework. Several HCI design approaches exist for cancer prediction [1]. The technique for predictive modeling we propose integrates HCI principles and includes an evaluation of the user interface for digital health systems [2]. While negotiation research encompasses many applications of cognitive HCI [3], the approaches for examining and assessing interacting patterns have been widely studied. The conventional categorization of affective, behavioral, and cognitive dimensions into positive, negative, and non-engagement introduces various fallacies [4].

This research investigates a number of machine learning strategies to perform the cancer categorization, identifying the most effective classifier based on parameters such as accuracy, precision, recall, specificity, F1 score, ROC AUC curve, training time, and maximum feature utilization.

The focus of this research is on five classifiers: the Decision Tree, the Random Forest, the SVM, XGBoosting, and ANN. We leverage the Wisconsin Breast Cancer (Diagnostic) dataset [5], comprising 569 samples and 30 features, with 212 malignant and 357 benign samples.

The purpose of this research is to detect breast cancer using classifier-based supervised machine learning methods that perform most efficiently on the given dataset and to decide the top predictive features. The remaining part of the article is structured into the following sections: Section 2 discusses the methods and findings of previous studies on breast cancer diagnosis. Section 3 discusses the planned technique for our study. Section 4 summarizes and examines the experimental outcomes. Finally, Section 5 concluded the article.

Determining the most precise and reliable algorithm for detecting breast cancer is the major goal of our work. The data analytic algorithm utilised for disease classification, clustering, anomalies and association [32]. To identify and classify breast tumors, we must access computer-assisted diagnostic data [33]. The proposed architecture shown below in the Figure 1. Here we assess the data and apply 5 supervised ML approaches to generate the correct decision using different parameter.

Imaging techniques such as X-ray, MRI, endoscope, ultrasound, and others produce good images but it harms the human body, so images are obtained with less energy, resulting in poor image quality and low contrast. Human computer interaction diagnosis (HCID) is a technique that improves image quality and interpreting mammography images to identify whether it is benign or malignant. Artificial intelligence diagnostic system, which measures the severity of the patient's symptoms and uses a membership function to address the patient's emotions [34]. Some proposed multi-feature disease fusion discrimination technique based on PD R-CNN [35]. As per research of the human breast cancer prognosis, prediction is performed utilizing diseased images and multiple kernel learning [36]. Proposed methodology for human computer interactive diagnosis:

1. Loading the data set.
2. Preprocess the data gathered through the digital image of the breast mass aspirated with a very little needle.
3. Apply machine learning techniques (DT, RF, SVM XGBoosting and ANN) on preprocessed dataset.
4. Find performance evaluation parameters (accuracy, recall, precision, specificity, F1 score, ROC AUC, training time and feature importance) for all classifier.
5. Find best classifier.

These five steps show the complete process of data classification and evaluation to find best classifier using evaluation parameter.

1.png

Figure 1. Proposed architecture

3.1 Dataset acquisition

2.png

Figure 2. Benign and malignant count

For this analysis, we utilized the WBCD dataset, which computes ten real-valued characteristics from each nucleus cell. The standard error and average of the three biggest results are also obtained for these ten attributes, giving each image a overall of thirty characteristics. The data set include 659 images with a class composition that includes 212 malignant and 357 benign as depicts in Figure 2.

3.png

Figure 3. Image analysis

The characteristics of the breast image are calculated using a digital image of a breast mass aspirate taken with a very small needle as shown in Figure 3 [37], this information utilized for assessing either a cancer is benign or malignant, we train the model to classify the dataset using above machine learning techniques.

Figure 3 provides a graphical visualization of a cell, showing the nucleus's long (dl) and short (ds) diameters together with the mathematical calculations for the nucleus's approximative area $\left(A=\frac{\pi d_l d_s}{4}\right)$ and elongation $\left(E=\frac{d_l}{d_s}\right)$ that were derived from those measure [37].

For each nucleus in this data set, the first ten core characteristics with real values are generated using ROI - radius (the mean distance between points on the perimeter and the center), texture (Grayscale value’s standard deviation) perimeter, area, smoothness (varying locally in radius), compactness (perimeter² / area - 1.0), concavity (severity of the contour's concave regions), concave points (concave regions there in the contour), symmetry, fractal dimension ("coastline approximation" - 1), We compare all ten features as shown in the Figure 4.

Twenty more features have been computed available in data set, ten features are based on mean and ten based on worst values of 10 core features, resulting 30 features for each image. All features describe the nucleus cell depicted in the image. X-ray mammography is used to obtain images.

4.png

Figure 4. Comparison of 10 core features

Early detection and medication are the current approaches to this disease. More patients must be discovered early in this disease to enhance survival. As a result, we evaluate current and new breast cancer screening and detection strategies to see where we might make improvements. This paper goes over some of the most regularly used breast cancer prediction supervised ML approaches and dives into emerging modifications [38]. As indicated in Table 1, 70 percent training and thirty percent testing.

Table 1. Training and test data set

We emphasize the comprehensive study of supervised ML approaches to classify the cancer is benign or malignant. We classify WBCD dataset using above five supervised ML approaches that will used for human computer interactive diagnosis. Some study based on identification of gene that is more correlated worth the prognosis of breast cancer [39]. IDS is implemented using CatBoost Classifier, it is a part of the ML ensemble strategy [40]. In order to improve breast cancer classification, Liew et al. [41] developed Deep Learning and Extreme Gradient Boosting, which divides binary breast cancer histology images into benign and malignant categories (DLXGB). In this era deep learning, mammography is considered computer-assisted detection as a use case for better human-computer interaction [42].

Benign and malignant dataset is to be classified using various machine learning algorithm. We have 10 basic features, each with a standard error and the average of the three greatest values, therefore data will be arranged into thirty features.

Human-computer interaction is vital for improving and maintaining high-level cancer detection skills. The results revealed that utilizing a low-cost display monitor and a variety of HCI approaches, observers may be trained to distinguish some critical breast cancer appearances. Deep learning applied to a specific clinical application for radiology image processing, which is known as computer-assisted diagnosis (CAD) [42]. To determine which features are helpful for classification, a box plot is used to show the classification. Since the median values for radius mean, perimeter mean, and area mean differ significantly in this instance, these characteristics are important for classifying cancer. However, the box plot of these characteristics is virtually same, thus we may select any feature from these features to decrease the size of the data collection or the computational complexity of the prediction as show in Figure 5. Box plot will show the data for all features having details of lower quartile (25% data below such value), upper quartile (25% data beyond this and 75% of data above this value), median (mid point of data), max denote highest value and min denote the lowest score.

5.1.png	5.2.png	5.3.png
5.4.png	5.5.png	5.6.png
5.7.png	5.8.png	5.9.png
5.10.png	5.11.png	5.12.png
5.13.png	5.14.png	5.15.png
5.16.png	5.17.png	5.18.png
5.19.png	5.20.png	5.21.png
5.22.png	5.23.png	5.24.png
5.25.png	5.26.png	5.27.png
5.28.png	5.29.png	5.30.png

Figure 5. Benign and malignant deviations in all 30 features

3.2 Supervised machine learning algorithms

Machine learning techniques applied for WBCD dataset are:

1. Decision tree classifiers provide a training model that predicts the category or value of the target variable using past data and basic decision criteria. The detailed approach to the decision tree, compare most used algorithm in decision tree, and found that decision approach had the highest accuracy in all other studies [43]. Decision tree is trained on default value of criterion that is gini which is the measure of purity and impurity while producing a decision tree using the CART technique. Gini index reduce the impurities through the root to the leafy nodes of a decision tree model.

2. A meta estimator called the random forest classifier fits a variety of decision trees to diverse dataset subsamples, improving prediction accuracy and reducing overfitting. A survey and future research directions, present the result for improvement in accuracy by using different split measures and combining functions [44]. The Number of trees are estimated with a default value 100. The working procedure can be outlined in the following steps:

1. Take K random points of data for the training set.
2. Build a decision tree using the selected data values.
3. Select the number N for the decision tree that we wish to build.
4. Repeat Step 1 & 2.
5. Determine each decision tree's estimates for newly data points and allocate them to the group with the most shares.

3. To efficiently categorize newly data points in future, support vector machines build the ideal decision boundary or line that separates an n-dimensional space into classes. A hyperplane is the optimal decision boundary, many applications built around SVM summarized challenges and trends, SVM has been utilized in numerous academic disciplines, including face recognition, text categorization, image classification, and handwritten character recognition, they find that SVM is not so popular when the data set is so large data set require huge training time and accuracy will become poor. So, it is enough for small data sets [45, 46]. Radial Basis Function (RBF) kernel is a prominent kernel function associated with SVM machine learning. It is a nonlinear kernel function that transfers the input data into a higher dimensional input domain using a Gaussian function.  The guassian kernel is expressed as $\exp \left(-\gamma\left\|x-x^{\prime}\right\|^2\right)$ where $\gamma$ is specified by parameter gamma, must be greater than 0. $\left\|\mathrm{x}-\mathrm{x}^{\prime}\right\|^2$ is squared Euclidean distance across the supplied vectors.

4. Extreme gradient boosting is a scalable machine learning approach that uses a distributed gradient-boosted decision tree. It offers concurrent tree boosting. It is a random forest-inspired decision tree ensemble learning algorithm. A power system transient stability prediction approach based on XGBoost makes predictions about stability depending on particular operating scenario features [47, 48]. The gradient boosting ensemble technique consist of three simple steps.

1. For predicting the desired variable y, an initial model F0 is created. This model will be associated with a residual (y minus F0).
2. The residuals from earlier stages are utilized to fit a new model, h1.
3. F0 and h1 are now combined to form F1, an enhanced version of F0. F1's mean squared error will be smaller than F0's:

$F_1(x)<-F_0(x)+h_1(x)$

To increase F1's performance, we could model after its residual and develop a new model F2:

$F_2(x)<-F_1(x)+h_2(x)$

This procedure can be performed m times until the residual is as small as desirable:

$F_m(x)<-F_{m-1}(x)+h_m(x)$

The active learner does not interfere with the preceding steps' functions, Individuals instead provide their own information to reduce inaccuracies.

5. A computing network called an artificial neural network is frequently modelled on the biological neural networks that give the structure of human brain. It has neurons that are interconnected through several network levels. The most recent developments are formed of biological neurons, a network of linked nodes that resemble the neurons in a biological brain. Neural networks are quite helpful when working with enormous data sets [29, 49]. To predict the breast cancer artificial neural network has been trained with in 100 epoch with 32 batch size. The model has been compile using adam optimizer and binary_crossentropy loss function.

We find the number of features utilized for each machine learning technique. We have 30 features, sorted index is calculated for each feature used for all five machine learning techniques. We extract 5 most usable features for machine learning technique.